Research map / Large language models / Instruction and fine-tuning
Learning Rewrite-Invariant Reasoning with Targeted Alternation Training
Mousa Arraf, Ido Guy, Kira Radinsky
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add targeted alternation in-context demonstrations (200) | GPT-5.2 (xhigh) · HLE (text only) | solve rate (%): 35.4 -> 38.1 | Improved |
| generate meaning-preserving rewrites when building reasoning graphs (ablation) | targeted alternation training · Big-Math | accuracy (%): 66.7 -> 74.8 | Improved |
| generate meaning-preserving rewrites when building reasoning graphs (ablation) | targeted alternation training · DROP | accuracy (%): 86.1 -> 91.9 | Improved |
| generate meaning-preserving rewrites when building reasoning graphs (ablation) | targeted alternation training · MMLU-Pro | accuracy (%): 76.6 -> 78.0 | Improved |
| use contrastive success/failure pairs instead of positive-only trajectories (ablation) | targeted alternation training · DROP (GPT-4.1-mini) | accuracy (%): 90.1 -> 91.9 | Improved |
| answer-only supervised fine-tuning | DeepSeek-V2-Lite · MMLU-Pro | accuracy (%): 60.1 -> 53.8 | Hurt |
| answer-only supervised fine-tuning | GPT-4.1-mini, Phi-4, Llama-3.3-70B · MMLU-Pro | accuracy (%): 76.5 -> 76.7 (GPT-4.1-mini) | No clear effect |
| inference-time rewrite ensembling (majority vote over rewrites) | Phi-4 · DROP | accuracy (%): 76.8 -> 82.2 | Improved |
| inference-time rewrite ensembling | GPT-4.1-mini · DROP | accuracy (%): 85.9 -> 85.0 | Hurt |
| hard-example mining via logit-margin uncertainty on rewrites | Llama-3.3-70B-Instruct · MMLU-Pro | accuracy (%): 71.9 -> 69.7 | Hurt |
| use GNN instead of logistic regression for the boundary-crossing predictor (ablation) | boundary-crossing predictor · AUC across source models | AUC: 70.43 -> 80.17 (DeepSeek) | Improved |
| use GNN instead of graph transformer for the predictor (ablation) | boundary-crossing predictor · AUC across source models | AUC: 77.20 -> 83.83 (Phi-4) | Improved |
Closest papers by meaning
- Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.87)
- Efficient Post-Training Refinement of Latent Reasoning in Large Language Models (AAAI 2026 · similarity 0.87)
- Training Language Models to Reason Efficiently (NeurIPS 2025 · similarity 0.86)
- RewriteLM: An Instruction-Tuned Large Language Model for Text Rewriting (AAAI 2024 · similarity 0.86)
- Probing to Refine: Reinforcement Distillation of LLM Reasoners via Explanatory Inversion (ICLR 2026 · similarity 0.86)
- TAIA: Large Language Models are Out-of-Distribution Data Learners (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
